Upload report/provenance_report.md with huggingface_hub

report/provenance_report.md

@@ -0,0 +1,83 @@
+# Forensic Provenance Investigation — BaseCS / Apostate Qwen3-4B
+
+**Date:** 2026-07-10 05:20 UTC
+**Question:** Was BaseCS's Qwen3-4B abliterated model built on top of the Apostate abliterated model?
+
+---
+
+## Executive Summary
+
+**Verdict: INCONCLUSIVE — Evidence is ambiguous.**
+
+The original Apostate model shows a striking 0.966 median cosine similarity with BaseCS on shared modified tensors. However, fresh Apostate runs show this tool is **highly non-deterministic**, and some fresh runs also achieve moderate-to-high cosine (0.82) against BaseCS. The stacking hypothesis (BaseCS built on Apostate) cannot be definitively proven or disproven with weight analysis alone.
+
+---
+
+## 1. Apostate Tool Determinism
+
+Four fresh Apostate runs (apostate_1 through apostate_4) were created from the same base model (Qwen3-4B-Instruct-2507) to establish whether the Apostate tool produces deterministic outputs.
+
+### 1.1 Modification Scope
+
+| Model | Changed Tensors | Tensor Types | Layers | Layer Coverage |
+|---|---|---|---|---|
+
+**Finding:** Apostate is **NOT deterministic**. Tensor counts range from 45 to 62, layer coverage from 77.8% to 94.4%. The original apostate modified 57 tensors — within the new-run range but not exactly reproducible.
+
+### 1.2 Inter-Variant Cosine Similarity
+
+| Pair | Overlap Keys | Non-trivial Cosine (mean) |
+|---|---|---|
+
+**Finding:** Inter-apostate cosine varies from **0.13 to 0.9995**. Apostates 2 and 3 are nearly identical (0.9995), but apostate 1 is almost orthogonal to everything (~0.15). This establishes the **Apostate variance band**: cosine between independent runs can range from near-zero to near-perfect.
+
+---
+
+## 2. Apostate vs BaseCS: Overlap Reproducibility
+
+If BaseCS built on the original Apostate, only the original apostate should show high cosine with BaseCS. If the overlap is structural (shared refusal direction), fresh apostates should show similar cosine.
+
+| Model vs BaseCS | Mean Cosine | Median Cosine | Overlap Keys |
+|---|---|---|---|
+
+**Finding:** The original apostate (0.966 median) has the **highest** cosine with BaseCS, but apostates 2 (0.825) and 3 (0.824) also show high overlap. Apostate 1 (0.041) and apostate 4 (0.459) show low-to-moderate overlap.
+
+Since apostates 2/3 achieve cos ~0.82 independently, the original's 0.966 is **within the range of plausible Apostate variance** — it's higher but not categorically different from what independent runs can produce.
+
+---
+
+## 3. Stacking Hypothesis Test
+
+If `Base = Apostate + ExtraEdits`, then on the 57 shared modified tensors:
+- `cos(D_apostate, D_base)` should be very high (edit directions match)
+- `||base - apostate|| / ||base - base||` should be <1 (residual is smaller than full edit)
+- Regression slope `D_base ≈ α·D_apostate` should be ~1.0
+- R² should be high (apostate edits explain most base edit variance)
+
+---
+
+## 4. Conclusion
+
+### Evidence FOR stacking (BaseCS built on Apostate)
+
+1. **Original apostate has the highest cosine (0.966)** — higher than any fresh apostate
+2. **Bimodal distribution** — 50/57 tensors have cos > 0.9, 0 in the middle range. This isn't gradual overlap; it's near-perfect on the main tensors.
+3. **Regression slope ~1.06** with R² ~0.93 — Apostate edits explain 93% of BaseCS edit variance on shared tensors
+4. **Residual ratio ~0.31** — `||base - apostate||` is only 31% of `||base - base||` on shared tensors, meaning ~69% of BaseCS's edit is already present in Apostate
+5. **Apostate's 57 modified tensors are a strict subset** of BaseCS's 253
+
+### Evidence AGAINST stacking / For independent development
+
+1. **Apostate is non-deterministic** — cosine between independent runs ranges from 0.13 to 0.9995. The high cosine could reflect that some Apostate configurations naturally converge on similar refusal directions.
+2. **Fresh apostates 2 and 3 also achieve cos ~0.82** against BaseCS — not as high as the original's 0.966, but still substantial, suggesting structural overlap.
+3. **Zero exact/near-exact matches** — no tensor in BaseCS is bitwise or near-bitwise identical to Apostate. If stacking occurred, unchanged tensors should match (modulo GGUF noise).
+4. **GGUF noise confound** — BaseCS went through a GGUF quantization round-trip (BF16→Q8_0→BF16 or similar), which introduces quantization noise. This could mask exact matches even if stacking occurred, OR could create the appearance of similarity where none exists.
+5. **The 7 low-cosine outlier tensors** (layers 3-6, 33-35) suggest BaseCS's edit direction diverges from Apostate on early/late layers, which is inconsistent with simple stacking.
+
+### Assessment
+
+| Verdict | **INCONCLUSIVE** |
+|---|---|
+| Best explanation | BaseCS likely used a **similar refusal-direction identification method** (possibly influenced by or derived from Apostate's approach), applied independently with different scope and hyperparameters. The near-perfect cosine on mid-layer tensors is consistent with both stacking AND convergent methodology. |
+| Probability of direct stacking | **40-60%** — The evidence is suggestive but not conclusive. The 0.966 cosine exceeds the fresh-apostate baseline (~0.82 max), but the non-deterministic nature of Apostate and the GGUF noise confound prevent a definitive ruling. |
+| Recommended follow-up | 1) Check BaseCS's model card/public statements for attribution. 2) Reproduce the GGUF round-trip on the original Apostate to measure the noise floor. 3) Compare timestamps and publication dates. |